How can I perform the negative_sampling by the given nodes?

[songsong0425]

Hi, I have a simple question about the negative_sampling function.

When I have the example edge_index as below, I want to run the negative_sampling based on the nodes(i.e., 0, 1, 2, 4, 5, 8) in the original edges.

edge_index = torch.as_tensor([[0, 0, 1, 2],
                              [8, 1, 5, 4]])

But it returned unexpected results including unseen nodes.

negative_sampling(edge_index)
# tensor([[8, 3, 1, 8],
#        [4, 8, 8, 3]])

How can I get the random negative edges from the nodes in edge_index?

[rusty1s]

edge_index = torch.as_tensor([[0, 0, 1, 2],
                              [8, 1, 5, 4]])

means that you have 9 nodes in total, while node 3 and node 7 are isolated nodes. If you want to exclude them, you have to shrink your edge_index down to not include isolated nodes:

edge_index = torch.as_tensor([[0, 0, 1, 2],
                              [3, 1, 5, 4]])

[songsong0425]

Thank you for your comment. But if the numbers in edge_index indicate the node ID, is it impossible to do negative sampling? (i.e., in the example, 3,6,7 don't have links) Currently, I made a dictionary for the existing nodes and performed negative sampling on the dictionary value then remapped to the key. Are there any efficient ways?

[rusty1s]

If examples 3, 6, 7 don't have links, then why shouldn't negative sampling sample negative links for these nodes? If you want to completely exclude nodes from negative sampling, it is best to fully discard them as node IDs IMO.

[songsong0425]

Sorry for the unclear explanation. Before negative_sampling, I performed a custom ID converting process across the original dataset, changing the node name to the unique IDs. Therefore, some nodes in the positive sets can be absent in the negative set, and vice versa. In this case, I want to perform sampling on the nodes in negative samples and unique nodes in the positive edges for the transductive setting to supplement fewer negatives. Please tell me if you need more explanation or if my approach is wrong.

[rusty1s]

If you have specific constraints on how negatives should look like (e.g., a specific subset of allowed negatives), then I think the most straightforward way is to simply write a custom negative_sampling function that fits your needs.

[songsong0425]

I see. Thank you for your suggestion. I'll try it.